Document AGILLM3.5 runtime code

README.md

@@ -20,6 +20,15 @@ Single full file per snapshot (each round is a block merge, not a delta).
 Checkpoint dict keys: `core` (backbone), `ar`, `sat` (heads), `cfg`, embedded
 `tokenizer_json`, plus `disagg_updates` (merge provenance) on the distributed master.
 
+
+## Code
+
+- `agillm35.py` - single-file AGILLM3.5 runtime for training/status/inference.
+- `distributed/public_join/` - public signed-lease host and outbound worker scripts for untrusted joiners.
+- `distributed/inference/agillm35_distributed_infer.py` - phase-1 distributed AR inference harness for transformer/MoE/DiffusionBlock layer stages.
+
 ## Inference
 Load with the AGILLM nB300 code (`infer --mode ar|sat`); the tokenizer round-trips from the
 embedded `tokenizer_json`.
+Distributed AR inference can split contiguous transformer/DiffusionBlock layer ranges across local and HTTP worker stages. The network payload path uses a raw tensor wire format rather than unpickling remote worker responses; use TLS and a bearer token outside localhost.
+